2025年12月5日TOP2在北京与一位常年做GPU优化的群友线下交流，其认为理想能让Orin跑VLA是一 个很有水准的工作。 认同"理想是在教英伟达怎么压榨芯片能力"这句话既偏颇又有事实依据是get到理想这项工作很有水 准的核心锚点。 能够显著优化寄存器复用效率（Register Reuse），有效规避Orin架构上因寄存器压力过大导致的溢出 （Register Spilling）。 理想可能通过显式插入Ampere架构特有的异步拷贝指令（Async Copy），将计算流水线与显存访问 的延迟掩盖（Latency Hiding）做到极致。 敢于投入资源深耕此领域，证明了理想团队具备深入分析SASS（流多处理器汇编）并在指令级挖掘 硬件潜能的核心能力。这是一项高门槛的系统工程。 2025年8月8日，理想詹锟表示："我们与英伟达进行了深度架构对齐，基于PTX（Parallel Thread Execution）指令集重构了底层算子，并打造了自研推理引擎，成功突破了通用TensorRT算子集的性能 天花板。" 詹锟提及的与英伟达深度交流，实质是指在Orin芯片（Ampere架构）微架构层面获得了原厂级的技 术指引 ...

2025年11月13日理想基座模型团队发布 模型在不同语言下的指令遵循能力存在显著差异，GPT-o3-2025-04-16中文任务的严格准确率为 76.9%，而英文任务为 63.5% (中文英文没打 错)。即单一语言的评测无法代表模型的整体控制能力 。 LexInstructEval(基于形式化语法的指令评估框架) 框架由两个解耦的工程模块组成: 数据构建管道 (Data Construction): 基于形式化语法生成无歧义的指令数据集 。 自动化验证引擎 (Automated Verification Engine): *基于确定性代码(而非 AI 模型)的验证工具 。 数据策略: 形式化语法分解 (Formal Grammar) 不直接使用自然语言生成数据，定义了一套形式化语法，将所有指令解构为机器可读的 (Procedure, Relation, Value) 三元组: LexInstructEval: Lexical Instruction Following Evaluation for Large Language Models 理想这篇论文核心指出限制模型性能的瓶颈不在于指令的数量(Bread ...

Core Viewpoint - The article discusses the recent developments and discussions surrounding the Ideal AI glasses, highlighting the need for further analysis and information gathering due to the rapid influx of news in the AI sector [1]. Group 1: Discussion Angles - The practicality of the first generation of Ideal AI glasses and the target audience for these products [2] - Perspectives from Li Xiang and other Ideal representatives on the long-term significance of AI glasses, including direct quotes from original materials [2] - Future iteration directions for Ideal glasses, comparison with mainstream industry players, and the product standards relative to competitors, including perceived advantages and disadvantages [2] - The potential market size for AI glasses, the necessity for Ideal to enter this market, and the appropriateness of the current timing for entry, as well as the relevance of AI glasses for automotive companies aiming for long-term survival [2] - Sales expectations for the first generation of Ideal glasses and the current status of these sales [2] - Collaboration between Ideal and Zeiss, including considerations for overseas markets and ambitions for the next 1-2 years [2]

2025年11月24日理想发布Compressor-VLA与AVA-VLA。 Compressor-VLA提出了一种针对机器人操作场景的高效视觉压缩方案，旨在解决端到端模型太重、太慢的落地难题。教会了机器人带着目的去观察，通 过语言指令过滤掉视觉垃圾，用更少的算力实现更精准的操作。 现在的具身智能大模型存在极其严重的算力浪费: 像一个强迫症患者，花费巨量算力去处理背景墙纸、地板纹理等无关紧要的视觉信息，导致推理延迟 过高，无法满足机器人的实时控制需求。 反直觉现象: 为了给模型减肥，传统做法是直接丢弃视觉token。这种盲目剪枝会导致灾难性后果——模型为了降低计算量，可能会保留纹理清晰的桌布 图案，丢弃了模糊但至关重要的物体边缘或把手位置。 现有的压缩算法是任务无关(Task-Agnostic)的。也就是说，压缩器是个瞎子，它只看图片本身，根本不知道机器人当下的任务是拿苹果还是关抽屉。这导 致在压缩过程中，关键的任务线索被当做噪声误删了 。 AVA-VLA针对现有端到端机器人大模型最核心的健忘问题，提出了一套工程化解决方案。 以下为更细化论述： 解决方案框架 采用了一种双通道互补结构，类似于指挥官 + 工匠的 ...

理想第二个十年，三大关键选择： 李想25Q3电话会议画饼压缩版： 组织： 从25Q4回归创业公司管理模式，放弃过去三年采用的职业经理人模式。 职业经理人模式成功三要素：1.行业和技术周期相对稳定。2.企业的行业地位领先且稳固。3.创始人 和创业团队没动力或不在了。案例：苹果/微软 创业公司模式成功三要素：1.行业和技术周期在发生巨大的变化。 2.行业格局不确定，企业还不是领 先者。 3.创始人和创业团队每天还在投入的努力工作，充满动力。案例：英伟达/特斯拉 产品： 停留在电动车，竞争逻辑只是参数与功能堆砌。未来10年具身智能最有价值的产品，一定是 具备自动和主动能力的汽车，提供飞机头等舱的服务与母亲般无微不至的照顾。 技术： 数字世界AI与物理AI是完全不同的系统能力。 具身智能系统包含：感知（眼睛、耳朵）模型（大脑）操作系统（神经）算力（心脏）本体（身 体）。行业急切需要改变的很多： 1.感知：行业现在的3D有效感知距离（非理论最大）为100多米，如果升级成人眼工作原理相似的3D 传感器，有效距离可以扩大2~3倍，今天辅助驾驶50%以上的常见问题将解决。3D感知不仅限于自动 驾驶的领域，在车外和车主的主动交 ...

Core Viewpoint - The article focuses on the evolution of Li Xiang's communication style with investors, highlighting a shift from a defensive posture to a more collaborative and constructive approach in discussing the company's vision and future. Group 1: Communication Evolution - In the April 2024 L6 launch, Li Xiang expressed frustration over being misunderstood by experts and the media, emphasizing that he aims to build cars for users rather than for the capital market [1] - By the November 2025 Q3 conference call, Li Xiang demonstrated a willingness to engage deeply with investors, spending nearly 15 minutes sharing his thoughts on the company's future and expressing gratitude for their support during challenging times [2] - The change in Li Xiang's communication reflects a transition from viewing investors as outsiders to recognizing them as partners in achieving a shared vision, indicating a strategic shift in how the company interacts with its stakeholders [3] Group 2: Strategic Intent - Li Xiang's remarks suggest a commitment to making Li Auto a leader in the field of embodied intelligence, aiming to provide high user value over the next 3 to 5 years [2] - The emphasis on maintaining the company's essence while adapting communication strategies indicates a desire to coexist with the market without compromising core values [2] - Li Xiang's reflections on refining his approach suggest a focus on effective communication that resonates with investors, aiming to align their understanding with the company's long-term goals [3]

2025年11月29日理想整车电动产品负责人唐华寅表示已经有国际大牌碳化硅芯片供应商和理想研发团 队联系获取技术IP和合作机会。 理想碳化硅芯片 用户价值： 1.自研电机/碳化硅电控系统/碳化硅芯片带来30公里续航提升。 2.支持整车高压网络从400V平台向更高的800V电压平台转变，达成更快的充电速度。 3.更快的开关速度支持了更精细的电机电流控制，降低了电磁噪声，提升了驾乘静谧性 。 4.碳化硅功率模块像车辆动力系统的心脏，通过它，电驱动系统将电池提供的直流电转变为电机所需 的交流电，从而输出动力。 行业因电流能力不足，为了降低通道电阻，选择浅PN结设计，导致极限工况下脆弱 。得益于电流能 力的富余，理想采用2倍能量注入的超深PN结，面对1000安培短路电流也能稳健夹断，测试通过率超 99.9% 。 行业为了压榨性能，不得不追求极致的光刻线宽，导致工艺容错率极低 。理想依靠元胞架 构的高通道优势，不再依赖极限线宽，将工艺制造窗口扩大了50%，显著提升了可制造性与良品率 。 5.耐压高、导热好、损耗低，显著减少电驱动系统的重量、体积。（从而有可以把省下来的重量体积 给其他配置） 6.基于长期大量测试数据，较 ...

Core Viewpoint - The article discusses the strategic choice of Yi Yang Qianxi as the brand ambassador for Li Auto, emphasizing the importance of taste and differentiation in branding and identity [1][2]. Group 1: Insights on Taste and Branding - The essence of taste is described as a tool for self-differentiation and identity signaling, serving social distinction needs [1]. - The pursuit of differentiation is a fundamental human instinct, akin to how products need differentiation to stand out in the market [1]. Group 2: Key References Supporting the Choice - In a conversation on "Hot Waves," Li Xiang took time to consider Yi Yang Qianxi as a suitable actor, highlighting his versatility and ability to embody different roles effectively [1]. - A peer in the AI industry noted that Yi Yang Qianxi has a unique appeal due to his consistent quality in film roles and his participation in a street dance variety show, which adds to his image of authenticity [2]. - A dialogue featuring Luo Yonghao and Bi Gan emphasized the high regard for Bi Gan's artistic expression, with Yi Yang Qianxi starring in Bi Gan's acclaimed film "Wild Times," indicating a strong connection between quality filmmaking and the ambassador's image [3]. Group 3: Importance of Communication Skills - The article highlights the high value placed on communication skills, suggesting that individuals with strong expression capabilities are often of high caliber [4].

Core Viewpoint - The article discusses the advancements in autonomous driving through the introduction of the AD-R1 framework, which utilizes closed-loop reinforcement learning to enhance safety and robustness in end-to-end autonomous driving systems, addressing the limitations of existing world models in predicting dangerous outcomes [2][4]. Group 1: Closed-Loop vs. Open-Loop Systems - Open-loop systems rely on offline data and static playback, while closed-loop systems interact dynamically with the environment, allowing for real-time adjustments to the vehicle's trajectory [1]. - The AD-R1 framework represents a significant step in closed-loop reinforcement learning for autonomous driving [1]. Group 2: Challenges in Imitation Learning - Imitation learning faces two main challenges: distribution shift due to unseen long-tail scenarios in the real world and the lack of negative feedback, making it difficult for AI to learn from mistakes [3]. - Optimistic bias is identified as a systemic flaw in reinforcement learning for autonomous driving, where models may generate unrealistic safe scenarios despite unsafe actions [3]. Group 3: AD-R1 Framework Components - The AD-R1 framework includes two core components: the development of an impartial world model and reinforcement learning based on future imaginings [4]. - The impartial world model employs counterfactual data synthesis to teach the model the consequences of unsafe driving behaviors [4]. Group 4: Model Training and Evaluation - The training process involves sampling candidate trajectories, imagining future scenarios using the impartial world model, scoring based on predicted outcomes, and updating the policy using the GRPO algorithm [8]. - The framework allows for detailed reward calculations through the use of 3D/4D voxel outputs, enhancing the evaluation of collision severity and ensuring vehicle stability on the road [8]. Group 5: Additional Features - Trajectory-aware gating is implemented to ensure the model focuses on relevant features along the driving path, while ego-trajectory fidelity loss penalizes deviations from the input control commands [6]. - The framework also includes volume collision penalties and vertical clearance checks to enhance safety in complex environments [8].

Core Viewpoint - The article discusses Li Auto's strategic choices for the next decade, focusing on organization, product, and technology as key areas for development and adaptation to market changes [2][9]. Group 1: Organization - Li Auto has transitioned from a startup management model to a professional management model over the past three years, recognizing the differences in governance systems between the two [2][6]. - The company believes that the current environment, characterized by rapid technological changes and industry uncertainty, aligns more with a startup management approach, emphasizing dialogue, user value, efficiency, and problem-solving [4][5]. Group 2: Product - The company is contemplating its product direction for the next decade, questioning whether to focus solely on electric vehicles, smart terminals, or embodied intelligence [6][8]. - A shift towards embodied intelligence is proposed, where vehicles could evolve into intelligent robots capable of providing automated and proactive services, enhancing user experience significantly [8][12]. Group 3: Technology - Li Auto emphasizes the need for a complete system technology choice, focusing on embodied intelligence rather than just language intelligence, which requires a different set of capabilities [9][10]. - The company is developing a comprehensive system for embodied intelligence, including perception, modeling, operating systems, computing power, and hardware modifications to enhance vehicle capabilities [11][12]. - The M100 chip is central to the company's future technology strategy, expected to significantly improve performance and cost-effectiveness in the next generation of products [29]. Group 4: Market Strategy and Financials - Li Auto is addressing supply chain challenges and production ramp-up for its i6 and i8 models, with plans to enhance battery supply through dual sourcing [18]. - The company is optimistic about future sales growth despite potential market fluctuations due to changes in government subsidies and taxes, focusing on user value and technological advancements to maintain competitiveness [20][22]. - The company has made significant investments in core technologies related to electric drive, battery, and electronic control systems, aiming to provide a reliable user experience [16][17].